The compute cost of the most expensive AI training run will reach about $230 billion (one percent of 2021 US GDP) around 2040.
Assessment
Credible evidence or argument exists on multiple sides.
The claim originates in a 2023 Epoch AI analysis by Ben Cottier, which extrapolated the dollar cost of the largest machine-learning training runs and forecast that a single run would cost about $233 billion, one percent of 2021 US GDP, around 2040, with a wide stated uncertainty spanning 2033 to 2062. The extrapolation's ingredients hold up individually: the most expensive run in 2025 cost on the order of several hundred million dollars, and costs for the largest runs have grown steadily since 2015, a pace that remained multi-fold per year through the mid-2020s.
What the forecast turns on is whether that growth continues through the 2030s, and this is genuinely disputed. Sustaining the trend to 2040 would require developers to spend tens and then hundreds of billions of dollars on individual runs, which critics argue is financially unsustainable given the gap between AI infrastructure spending and AI revenues; a related prediction that spending growth will slow substantially during the 2020s is itself contested. On the other side, the trend has repeatedly outlasted predictions of its end, and feasibility analyses find no hard technical barrier through at least 2030.
The date could also miss in the other direction: costs through the mid-2020s have grown faster than the rate the forecast extrapolates, and at that faster pace the milestone would arrive in the early-to-mid 2030s rather than around 2040. The question resolves progressively as the decade unfolds; a clear break or persistence in frontier training budgets through the late 2020s would substantially narrow the range.
Full reasoning: the evidence and decisions behind this verdict
The sole recorded source instance is the originating forecast itself: Epoch AI's "Trends in the dollar training cost of machine learning systems" (Cottier, 2023, epoch.ai/blog/trends-in-the-dollar-training-cost-of-machine-learning-systems), which extrapolates from a 2025 base cost of $380M (90% CI $55M to $1.5B) at 0.2 orders of magnitude per year (90% CI 0.1 to 0.3) to reach $233.2B in 2040 (90% CI 2033 to 2062). The author explicitly frames $233B as one percent of 2021 US GDP in real terms, following Cotra's use of that threshold as a marker of maximal societal investment in a single run.
The base point checks out against independent estimates: Stanford's AI Index figures of roughly $78-100M for GPT-4 and about $190M for Gemini Ultra, with Epoch's separate cost study (epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models) projecting billion-dollar runs by 2027, are consistent with a several-hundred-million-dollar most-expensive run in 2025. The historical growth premise is likewise well anchored.
The verdict rests on the continuation premise, that rapid cost growth persists through the 2030s, which is the live crux. Against it: the argument that multi-fold spending growth is financially unsustainable through 2030, resting on the widely cited gap between AI infrastructure spending and AI revenues, and the related slowdown prediction, that spending growth slows substantially during the 2020s, which stands assessed as contested with the burden so far shifting toward slowdown proponents as budgets kept growing. For it: the trend's decade-plus persistence, and Epoch's feasibility analysis (epoch.ai/blog/can-ai-scaling-continue-through-2030) concluding runs around 2e29 FLOP are likely possible by 2030 given power, chip, data, and latency constraints, so any break before then would come from investment choices rather than hard limits.
Two failure directions keep the credence low even granting substantial probability to continued scaling. If growth persists at the 2 to 3x per year pace actually observed through the mid-2020s, the $233B level arrives around the early-to-mid 2030s, making "around 2040" wrong by being late; if the trend breaks for financial reasons, the level arrives much later or never, making it wrong by being early. "Around 2040" is the median of a distribution whose own 90% interval spans nearly three decades, so a credence of roughly 0.25 for the milestone landing in the neighborhood of 2040 (roughly the mid-2030s to mid-2040s) is an honest summary. Contested is the right status because the load-bearing premise divides credible parties; the claim is not merely unsupported, since the extrapolation is real evidence, nor supported, since the affirming evidence is a single organization's admittedly wide extrapolation and the discourse credibly disputes its central premise. What would change the conclusion: frontier training budgets through 2028-2030 either breaking sharply (toward contradicted) or continuing multi-fold growth with revenues to match (toward supported, though likely with an earlier date than 2040).
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because the most expensive training run in 2025 cost several hundred million dollars and costs for the largest runs have grown about 0.2 orders of magnitude per year since 2015, a growth rate that remained multi-fold through the mid-2020s, then if that growth continues through the 2030s, extrapolation reaches roughly $230 billion, about one percent of 2021 US GDP, around 2040.
Granting its premises the arithmetic goes through, but only at the extrapolation's own best-guess rate: the argument lives or dies on continued rapid cost growth through the 2030s, which remains the disputed crux, while the base point and historical trend premises are well evidenced. A further caveat cuts against the date itself: the multi-fold growth observed through the mid-2020s runs faster than the extrapolated rate, and if it persisted the milestone would arrive in the early-to-mid 2030s rather than around 2040.
Because growth in AI training spending at recent multi-fold annual rates is argued to be financially unsustainable through 2030, and spending growth on the largest runs is predicted to slow substantially during the 2020s, the extrapolated path would break more than a decade before a single run reaches roughly $230 billion, pushing the milestone well past 2040 or preventing it altogether.
If either premise holds, the conclusion follows: a trend that breaks in the late 2020s cannot deliver a roughly $230 billion run around 2040 by extrapolation. Both premises are themselves unsettled; the predicted 2020s slowdown stands contested, with spending growth so far persisting rather than slowing, and the financial unsustainability argument rests on a spending-revenue gap that continued AI revenue growth could close. The argument's force therefore tracks how those premises resolve toward the end of the decade.
Provenance
Where this claim has been said, linked to its canonical form.
extrapolating from the year 2025 with an initial cost of $380M (90% CI: $55M to $1.5B) using a growth rate of 0.2 OOMs/year (90% CI: 0.1 to 0.3 OOMs/year), a cost of $233.2B would be reached in the year 2040 (90% CI: 2033 to 2062)
All-things-considered view: extrapolating from the year 2025... a cost of $233.2B would be reached in the year 2040. This is the author's best-guess forecast of when training cost reaches 1% of US GDP.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Aug 11, 2026. Every judgment on this page is accompanied by a reasoning trace.